Inside Anthropic’s 2026 Developer Conference

Was this newsletter forwarded to you? Sign up to get it in your inbox.

To our surprise, the biggest launch from Anthropic’s developer conference in San Francisco yesterday wasn’t a model or a feature. Instead, it was the company’s announcement of a deal with SpaceX to allocate all of the capacity in the latter’s Colossus supercluster to Claude.

Anthropic has been riding a historic demand surge over the last year as Claude Code opened up a new wave of agentic coding for engineers and non-engineers alike. But compute constraints have caused friction even amongst its most die-hard fans—we’ve written previously about being frustrated with its OpenClaw restrictions and the speed of its latest models like Opus 4.7.

The deal with SpaceX changes that equation. Anthropic has already doubled rate limits for subscription plans, removed peak-hour limits on Pro and Max accounts, and raised API rate limits by as much as almost 17 times for certain tiers.

Other than that, the big story is Claude Managed Agents, Anthropic’s hosted agent product. The company released three new features:

• Multi-agent orchestration: a coordinator agent that spins up subagents in parallel baked into the platform
• Dreaming: Anthropic’s general-purpose version of compound engineering, a feature that allows agents to learn from past sessions to improve between runs
• Outcomes: Anthropic’s answer to Codex’s /goals command, allowing developers to specify an outcome and run an agent in a loop until the outcome is achieved

By themselves, these features are nice but not groundbreaking. What’s more important is that what an AI platform is has changed. In the GPT-3 days, the platform was a text completion end-point: Send text in, get text out. Now, with Claude Managed Agents, the platform is an AI model with a harness and host computer—all provided with unlimited scaling by the model companies.

Cora general manager Kieran Klaassen and I reported live from conference with our biggest takeaways, including the xAI compute deal, doubled Claude usage limits, Claude Managed Agents, and why the battle lines between OpenAI and Anthropic are starting to become clearer. Watch now:

We also recorded a conversation with Angela Jiang, head of product for the Cloud Platform, and Katelyn Lesse, head of platform engineering. The full episode drops tomorrow on AI & I—highlights below.—Dan Shipper

Building blocks that help your agents compose elegant backends

Agents right now often write code that needs to be cleaned up because they need three things most infrastructure wasn’t built to provide: immediate feedback, local reasoning, and a thin meta layer. Convex provides immediate feedback, enables local reasoning, and minimizes the meta layer, giving agents a surface layer that’s actually small enough to work with. Scale without breaking or spending a fortune.

Try Convex

Want to sponsor Every? Click here.

Vibe Check: Claude Managed Agents

Spiral general manager Marcus Moretti uses the platform’s new features

Anthropic launched Claude Managed Agents in April, and since then, Every’s AI writing tool Spiral has used the platform to power its API and command line interface (CLI), which lets developers and other agents talk to Spiral outside the web app. Claude Managed Agents run on Anthropic’s servers, instead of us having to run them on our own.

We set up a new Managed Agent in an afternoon and deployed it to power our API the next day. We’ve incorporated two of the new features Anthropic announced yesterday (memory and multi-agent orchestration) and are deploying the third (outcomes) soon.

Memory: Every’s editorial and social expertise—how to write a good X post, for example—lives in an Anthropic-hosted global memory store. The memory store lets us avoid including every piece of editorial and social expertise in the agent system prompt—the standing instructions that tell the agent what to do every time it runs. When a user asks for a podcast description, the agent doesn’t need to also recall how to craft a great LinkedIn post. It only pulls the relevant expertise with each request, thereby making responses faster.

Each Spiral subscriber also gets their own personal memory store. When you tell Spiral that you prefer em-dashes over semicolons or that your company name is one word and not two, it will remember and apply your rules by default the next time you run it.

Multi-agent orchestration: When users request a single draft of a piece of writing, one agent using Opus 4.6 Fast handles the workflow end-to-end. For multi-draft requests, a coordinator agent using Haiku 4.5 spins up multiple Opus 4.6 Fast subagents to compose drafts in parallel. Before multiagent orchestration, multi-draft requests were handled serially, and each draft added 20 to 30 seconds to the overall request time. A multiagent approach also reduced our costs for multi-draft requests by about a third because we were able to use cheaper models for part of the work.

Outcomes: Anthropic’s new outcomes capability is a feedback loop where one “grader” AI checks another AI’s work against a specified goal. Spiral’s main value proposition is writing quality, so we’re using outcomes to set up a rubric to ensure the writer agent’s output meets Spiral’s editorial standards and matches the user’s style guide. The rubric the grader AI uses is generated on-the-fly based on the global standards, the user’s writing style, and their writing preferences from memory.

Memory and multi-agent orchestration are live in production, and outcomes is coming soon. You can see the features in action by running npm i -g @every-env/spiral-cli && spiral login or logging into Spiral and using the install command on the Agent and API keys page.

Having set these features up in production, here’s what I think:

You are not totally locked into Anthropic’s universe. Every engineer worries that when a company offers a hosted version of something, it will be hard to leave. With Managed Agents, the agents themselves, sessions, and memory are all stored on Anthropic machines, and the agents themselves can only be powered by Claude—a managed agent can’t run on GPT-5.5 or Gemini.

I’ve mitigated this lock-in in two ways: First, we save agent runs to our own database in addition to Anthropic’s. This way, chats from the API appear in the web app just as web chats do, but it doubles as a safety net. If we ever wanted to leave Anthropic, we’d have all our historical data. Second, the Managed Agents platform lets you define custom tools for the agents. Those tools run on our servers, which means we can use whatever model we want inside the tools themselves. The coordinator agent is locked to Claude, but we control the layer underneath.

Using multiple agents has trade-offs. Multi-agent orchestration has allowed us to create multiple drafts faster and cheaper. However, coordination between agents adds overhead that prevents greater speed gains. Debugging also gets harder: If a Spiral draft comes back subpar, we have to investigate both the coordinator agent and the writer agent to identify the root cause. I’d recommend multi-agent orchestration only when your agent benefits from running subagents in parallel or using a mixture of models. Otherwise, a single agent works well.

Memory’s design is intuitive. Each memory is just a folder of markdown files, and each memory store is attached to a session with instructions that tell the agent when to consult it. Anthropic designed this feature thoughtfully—they kept it simple.—Marcus Moretti

The feature to watch: Dreaming

Cora general manager Kieran Klaassen sees his own philosophy mirrored back at him

Kieran has spent the last year trying to get agents to learn his preferences instead of forcing him to restate them every time. That’s compound engineering in a nutshell—each run leaves the system better prepared for the next one. So when Anthropic officially announced dreaming at yesterday’s Code with Claude event, he had a familiar feeling: The thing he’d been building was now a feature.

Dreaming is Anthropic’s name for a background process that reviews an agent’s past sessions and memory stores, finds patterns, and rewrites memory so the agent improves between runs. OpenClaw introduced a similar feature in April, but Anthropic’s take seems more focused on what teams of agents learn collectively than what a single agent remembers. The system learns from repeated corrections, recurring mistakes, and workflows that run well—creating, over time, an institutional knowledge base.

The feature currently lives inside Claude Managed Agents as a research preview, which is where Marcus has been testing it—with early success. Every plans to have its production agents dream as soon as the feature ships in a stable public release. But Kieran’s immediate question was: When is this coming to Claude Code?

Claude Code, after all, is where developers spend their days teaching agents the same repo quirks, the same testing rituals, the same “please don’t do it that way” preferences. Those preferences can go into memory files, but memory files get messy. They collect duplicates, stale rules, one-off notes, and contradictions—and as Marcus notes, memory introduces overhead, so you trade speed for quality every time you use it.

A dream cleans that up. It takes up to 100 past sessions and produces a reorganized memory store with duplicates merged, contradicted entries replaced, and new insights pulled out—memory that organizes itself, in Marcus’s framing. If Anthropic brings that loop to Claude Code, memory starts to look less like a notes folder and more like accumulated taste.—Katie Parrott

Inside Anthropic

What the company’s platform team told us off-stage

While at the conference, Dan sat down with Angela Jiang, Anthropic’s head of product for the Cloud Platform, and Katelyn Lesse, head of platform engineering, for a recorded conversation. Three things that stood out:

The generic harness is dead. Angela told us that building a generalized harness that lets you switch any underlying model for a different one—standard practice even a few months ago—is a losing strategy. Different harnesses paired with the same model produce “drastically different” results on Anthropic’s own evaluations. When the team built memory for Managed Agents, they tested multiple harness designs, and the performance gaps were large enough to make model selection feel secondary.

Our own experience backs this up: Our agents run on Claude with a harness tuned specifically for how Claude works. If we don’t want to risk getting locked in, we have to—as Marcus writes above—build the harness in a way that lets us swap in GPT or Gemini. But Angela’s argument is that the bigger risk is leaving performance on the table.

Infrastructure is the real wall. Katelyn told us that most people building agents expect the hard part to be the prompting, context window management, and tool setup required to get the most out of the model. In practice, everyone hits the same wall: infrastructure. They have to keep servers running, securely sandbox, prevent connection drops, and store transcripts. Before Marcus set up Managed Agents in an afternoon and deployed it the next day, we spent months on exactly that kind of plumbing.

Your agent needs a babysitter. Dan raised this problem directly: Agents get stale fast, running old models and old prompts with nobody responsible for updating them. Our solution so far has been to assign every agent an owner to keep an eye on it. Katelyn said the Anthropic team has built skills to help agents upgrade themselves to new models. “The most AGI-pilled people,” she added, “are running agents that monitor their agents.”

The full episode with Angela and Katelyn drops tomorrow on AI & I—we go deeper on where the platform is headed, what “outcome + budget” means as a design philosophy, and why Anthropic thinks Claude should eventually pick its own sub-agents.—KP